Skip to content

Decide Expert Talk context and request budgets #260

Description

@elkaix

Question

What deterministic context and provider-request budgets make the fixed Expert Talk protocol safe and predictable across two configured models?

Resolve the usable-context calculation for both participants and the fresh Fusion Lead, maximum output and tool-result budgets for each stage, maximum model steps and provider retries per stage, actual-usage reporting, and the admission test for complete peer-review and fusion packets.

The contract must preserve complete labeled stage outputs. It must not use silent truncation, provider-specific hidden compression, or a filesystem artifact as a missing-context side channel.

Use concrete scenarios: the smaller model has a narrow context window, one opening reaches its output cap, a read tool returns a large result, or the provider retries an otherwise valid stage.

Metadata

Metadata

Assignees

Labels

wayfinder:grillingWayfinder human decision conversation

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions