v0.9.34 — LLM intent classifier module + empirical bench
Added
- LLM intent classifier module (
agent::llm_classifier): fail-open fast-model classification with 5-second timeout and 20-token output cap so per-call cost stays trivial. 9 unit tests. - Empirical bench (
intent_classifier_bench_run_corpus,#[ignore]d): runs the classifier against a 27-case hand-curated corpus and reports agreement vs heuristic baseline plus latency.
Initial bench results
Tested on `google/gemini-2.5-flash` via OpenRouter:
| Metric | Result |
|---|---|
| Agreement with heuristic | 24/27 (88.9%) |
| LLM failures | 0 |
| Avg latency | 462 ms |
Disagreements were a mix: legitimate LLM wins on implicit fact-sharing the regex misses, one heuristic win on a compound recall+action task.
Conclusion
The classifier is a useful shadow signal but the latency is too high for synchronous primary use. A follow-up release will wire it as fire-and-forget shadow mode (off by default).
No behavior change in this release — module and bench only. All 2188 library tests pass.
Full Changelog: v0.9.33...v0.9.34