dsh-tool-bandit-search

LIMITATIONS

已知限制

- **Bandit state is in-memory** and resets on every restart. Persisting it via `ctx.storage` (which dsh already exposes) is a natural next step. - **Reward is a heuristic**, not a measure of actual answer quality — it captures result count and latency, not whether the results were *relevant* or *correct*. A stronger version might score reward against whether the model's final answer actually used the returned sources. - **The model can still issue multiple `search` calls per turn** even when `thorough` is already broadening internally — the plugin optimizes strategy *per call*, not the model's own multi-call behavior. - Built and tested against dsh's developer preview; the plugin/tool APIs may change before a stable release.