dsh-llm-cursor

LIMITATIONS

Known limitations

- HTTP/2 (ALPN) to `api2.cursor.sh` is required; there is no proxy bridge. - The CLI version pin can break when Cursor ships a new CLI that the pin no longer satisfies. Changelog that change when it happens. - Usage percents come from unofficial dashboard rails, not an official usage API. - Token usage chunks from `Run` do not include cache fields, so DSH cache-hit rate stays empty. - Fast SKUs are separate catalog families (`gpt-5.2` vs `gpt-5.2-fast`), not a third picker toggle. - 1M SKUs appear in Fetch for families Cursor offers Max Context (`claude-opus-5` vs `claude-opus-5-1m`). Saving does not re-insert a Max row you left unchecked. The Max row always sends `maxMode: true`. Composer and Cursor Grok do not get a 1M row. - You can also add a generic context row yourself (`claude-opus-5-272k`). The plugin peels a trailing `-<n>k` / `-<n>m` before talking to Cursor; DSH uses `n×1000` / `n×1,000,000` as the compaction budget. Cursor's API only has a binary `maxMode`, so a 272K row still sends `maxMode: false` — the suffix only changes DSH's compaction trigger. Product names such as `kimi-k3-max` are not treated as a context tier. The composer picker groups sibling rows that share a base id.