@bananiceee/dsh-zhushou

LIMITATIONS

Known limitations

1. **Top-P not wired yet** — DSH `GenerateOptions` / `LlmCallConfig` has no `topP` field today; the setting is kept, validated and persisted but not passed to `ctx.llm.stream`. It will be wired once the official API supports it. 2. **Session model UI does not follow activation** — persona/model/parameter overrides act on the `llm/stream` **request layer** (activation pipeline, see docs/ARCHITECTURE-ACTIVATION.md); the session header and DSH UI still show the original provider/model. 3. **The official system prompt can be replaced** — on activation, `rebuilt.system` is the assistant-composed block (persona first); if the assistant system prompt is empty the official system prompt is kept. 4. **Injections don't appear in the session log** — activation only rewrites the request face; `session/event` records the original messages (consistency first). 5. **reasoningEffort is mapped to the nearest tier** — DSH only has off/low/high/max; medium is mapped to low host-side, auto is omitted (server default). 6. **Memory extraction not closed-loop** — memory injection (keyword + time-decay top 5) and CRUD are implemented; the async "fast-model auto-extracts facts → memory pool" step is not wired yet (entries are maintained manually on the settings page). 7. **Skill panel reflects the host runtime** — `ctx.skills.list()` returns the skills registered in the current runtime (e.g. vision-skills), not a compile-time full catalog. 8. **No multimodal support** — text messages only; images etc. are planned for later versions.