What's new

• Continuation prompt state inheritance — short follow-up prompts (yes, ok, go ahead, do it, sounds good) instantly reuse the prior turn's classification instead of re-running the full Ollama/API classifier (~1–3s saved per continuation)
• Negative downgrades — no, stop, skip, cancel route to query/simple → llm_query
• Session-scoped state — persisted at ~/.llm-router/last_route_{session_id}.json with 30-min TTL; parallel sessions never interfere
• [via context-inherit] method tag visible in routing directive
• Hook bumped to v16

All changes since v2.2.0

| Version | Headline |
|---------|----------|
| v2.3.0 | Zero-Friction Activation — shadow/suggest/enforce modes, yearly projection, weekly digest |
| v2.4.0 | Repo-Aware YAML Config — .llm-router.yml, block_providers, model pins, config CLI |
| v2.5.0 | Context-Aware Routing — continuation prompts inherit prior route |

Upgrade

```bash
pip install --upgrade claude-code-llm-router
llm-router install

If using CC plugin:

claude plugin uninstall llm-router && claude plugin install llm-router
```