[Parked pending benchmark evidence] Intent-aware browse query planner

Context / Problem

• Weak browse queries reduce first-pass routing quality.
• Users repeat queries with trial-and-error wording.
• Clarification loops add latency and cost.

Proposal

• Add deterministic query planner for browse prompts.
• Suggest query variants before route execution.
• Surface reason tags for each suggested variant.
• Keep manual query entry always available.

Acceptance Criteria

• Planner produces stable suggestions for same input.
• Suggestions stay within defined token budget.
• Users can accept or skip suggestions.
• Benchmarks show improved first-pass selection rate.

Implementation Notes

• Keep planner rules explicit and editable.
• Do not auto-execute without user choice.
• Preserve existing browse path by default.

Tasks / Checklist

• Define planner policy and output schema.
• Implement suggestion generation and ranking.
• Add regression tests and benchmark cases.
• Document usage patterns and limits.

References

• Anthropic routing workflow guidance: https://www.anthropic.com/engineering/building-effective-agents
• MCP tools protocol model: https://modelcontextprotocol.io/specification/2025-06-18/server/tools

Start by reviewing the linked Anthropic routing guidance and MCP tools protocol, then define the planner policy and output schema described in the issue. Benchmark cases should establish stable suggestions, token-budget compliance, and improved first-pass selection before implementation; regression tests and usage documentation are also required.