[routing] Context-aware tool shortlisting with conversation hints

[dgenio]

Problem

Current routing uses only the query text for beam search scoring, ignoring conversation context. Multi-turn routing quality degrades because each query is scored in isolation.

Proposal

Add optional context_hints: list[str] | None parameter to Router.route():

result = router.route(
    query="send the report",
    context_hints=["email", "spreadsheet", "quarterly revenue"],
)

Compute TF-IDF similarity between hints and node descriptions; apply a configurable boost multiplier (default 0.15) with recency weighting. Include context_boost_applied in RouteResult for debuggability.

Constraints

• Backward-compatible: hints are optional
• Single-turn quality must not degrade (boost is additive)
• Deterministic behavior maintained

Acceptance Criteria

• Router.route() accepts optional context_hints parameter
• Context hints boost scores for matching tools
• No regression on single-turn queries
• RouteResult includes context boost metadata
• Boost multiplier is configurable
• Tests cover hint boosting and no-hint baseline

File Paths

• src/contextweaver/routing/router.py (edit)
• tests/test_router.py (edit)

References

• HuggingGPT model selection: https://arxiv.org/abs/2303.17580
• Agent memory: https://lilianweng.github.io/posts/2023-06-23-agent/