Durable Research and Visible Execution

Durable Research and Visible Execution

We hardened long-running research so users can understand what is happening, what evidence supports the result, and how to recover when an external source or worker is unavailable.

Progress users can trust

• Added persisted, user-safe execution progress with meaningful stages instead of generic waiting indicators.
• Added live event delivery, heartbeats, recovery states, and reconnect handling for active research runs.
• Improved the handoff between planning, source collection, synthesis, and report publication.

Evidence-grounded results

• Research findings and citations now stay connected to accepted source evidence.
• Improved source metadata, citation rendering, confidence context, and partial-result handling.
• Added clearer reporting when a source is unavailable or a result cannot yet be confirmed.

Safer and more reliable execution

• Hardened retries, dispatch, worker communication, and settlement around long-running work.
• Improved idempotency and failure recovery so a retry is less likely to duplicate work or charges.
• Added production inspection and smoke-test paths for the research execution lifecycle.

Research now behaves more like an inspectable piece of work and less like an opaque chat response.