@mars-sea/dsh-commandcode-provider

LIMITATIONS

Known limitations

- **Image input is model-gated** — only Vision models accept images; text-only models refuse them. - Switching to a text-only model in an image-bearing session is rejected by dsh — pick a model marked *`Image`* or remove the images first. - **Long image-heavy sessions keep working** — the provider caps the size of a single request (around 50 MB, and it is undocumented), so a session that accumulates many screenshots would otherwise start failing every request once it grew past that. The oldest images beyond the budget are replaced by a short "image omitted" note that names the attachment, the newest ones are sent as usual, and a request the provider still rejects as too large is retried once with a much smaller image budget before failing. - **No `stop` sequences** — requests carrying one fail. - Historical reasoning is replayed on both transports so a tool-calling loop keeps its chain of thought: `/provider/v1/chat/completions` passes it back as `reasoning_content`, and `/alpha/generate` carries it as a `reasoning` block inside the rebuilt assistant turn (the shape the official CLI sends) — a DeepSeek thinking-mode tool loop is rejected without it. Only tool calls with a paired tool result are replayed on either transport. - Every tool's parameter schema is normalized to the object root the provider requires, so a tool declared without one (a third-party plugin's hand-written schema, an MCP server's own `inputSchema`, a generated root `$ref`) no longer fails the whole turn with `schema must be a JSON Schema of 'type: "object"'`. - The model catalog is browsable without a key; chat requests need one.