Compaction evidence does not settle effects

Context

The default history evidence policy omits failed assistants. Compaction selects a contiguous complete prefix, so an omitted failed turn also protects completed turns after it. The existing explicit assistant-marked policy admits that failed evidence once the whole tool chronology is complete. A summarizer receives raw records; a failure that has no provider content must still be identifiable when those records are projected.

A matching error tool result can close the provider exchange while its external effect remains uncertain. Treating those two facts as one would either repeat a side effect or block compaction permanently for a receipt the canonical journal already retains.

Decision

Keep the existing selector and explicit shared evidence policy. Default omission remains a barrier. Every opted-in failed assistant, including an empty one, projects a non-success marker. The summarizer owns its schema and prompt and must preserve failed and uncertain outcomes rather than promote partial evidence to success. A record that cannot be safely projected is a reason to refuse that summary, not to delete the record from its range.

Compaction replaces one exact contiguous history range under snapshot CAS; it does not erase canonical events, effect proofs or retained terminal admission receipts. Effect recovery keeps using its original conversation/run/name identity: accepted returns the proof, intent reconciles, uncertain never repeats the effect. A new identity is new work, and a standalone caller still owns its exclusive execution lease.

Consequences

Applications can explicitly compact settled failed turns without rewriting statuses, inventing another selector or losing effect recovery. Incomplete tool exchanges remain barriers. Provider-valid chronology is not proof that an external operation succeeded, failed or is safe to retry. The model’s summary never becomes execution authority.