Reliable bulk work at scale

Six changes shipped in this release. Bulk work stopped depending on the assistant’s conversation window, large call histories stopped blocking the interface, and prompts were reorganized so most of each turn is served from cache.

Bulk operations

• Durable progress — asking the assistant to work through a large set of contacts now writes progress to a scratchpad that survives beyond the conversation window.
• Resume, never repeat — if context resets mid-operation it continues exactly where it stopped, with no duplicate messages and no skipped contacts.
• Batch reads — new tools fetch enriched contact data in batches, including memory snippets and message previews.
• Safe at scale — organization scoping, deduplication, and per-contact status tracking keep a long operation clean.

Call history sync

• Returns immediately — triggering a sync no longer waits for the whole download.
• A page at a time — calls are fetched and stored incrementally with live progress and bounded memory regardless of volume.
• No duplicate processing — a lease system prevents two workers from handling the same recording and safely reclaims stalled work.

Performance

• Prompts restructured — static content first and dynamic context last, so most of the prompt is reused between turns.
• Cheaper and quicker turns — roughly four in five prompt segments are now served from cache, with several hundred redundant tokens removed without losing capability.