Continue Deep Research

Treat the user's research as an evolving evidence base. Preserve useful prior work, challenge unsupported claims, and return the verified research delta rather than a rewritten summary of the same material.

Adapt to the available runtime

Do not refuse research solely because a preferred model, reasoning level, or subagent type is unavailable. Establish the available capabilities before planning the work, then preserve the evidence standard with the best supported execution shape:

1. Use the current root agent to ingest the evidence base, verify the highest-impact claims, and integrate the delta.
2. Use independent research lanes only when subagents are available, permitted, and materially improve coverage. Select only models and reasoning levels explicitly exposed by the runtime.
3. If subagents are unavailable or inappropriate, perform a bounded root-only audit. Reduce the number of open questions if necessary rather than inventing coverage.
4. State the execution shape, incomplete lanes, and any coverage trade-off in the closeout. A missing capability is a limitation to disclose, not a reason to abandon a well-scoped request.

Never claim that a model identity or capability was verified when it was not. Never fabricate an unsupported model selector.

Preserve the research boundary