- Providers get a default reasoning effort (low..max, empty = not sent);
chat can override it per message. OpenAI sends reasoning_effort,
Claude sends output_config.effort with adaptive thinking, and thinking
blocks are replayed during the tool loop.
- Show a thinking indicator while waiting for model output.
- Surface in-stream upstream errors (e.g. overloaded) instead of
breaking the SSE stream, and retry them up to 3 times when nothing
has been emitted yet.
- Model providers for OpenAI and Anthropic message formats (e.g. via
Sub2API), with API keys encrypted at rest and never returned
- Skills tree with markdown editing, priorities, drafts, publishing and
rollback; only published skills enter the chat prompt
- Chat page with tool calls over collected news and Binance USD-M
klines, a persisted CONTEXT.md, reset, and stop-generation support
- Chat may only create skill drafts and update CONTEXT.md; publishing,
deleting, moving and priority changes stay with the user
Let the crawler console set a backfill target per source. Each scheduled
run pages backwards from a checkpoint for at most 20s so real-time
crawling is not blocked. The checkpoint is saved only after a successful
run and is also kept in process memory, because the next trigger may be
queued before the database write lands. cls_depth is excluded since its
list API cannot page into history.
Seed um BTCUSDT/ETHUSDT/BNBUSDT as default symbols. The legacy futures
route fstream.binance.com/stream accepts connections but no longer pushes
data, so switch to /market/stream. Receive with a timeout so symbol
changes are picked up and silent connections are reconnected.
eezd