Skip to content

Latest commit

 

History

History
24 lines (20 loc) · 1.23 KB

File metadata and controls

24 lines (20 loc) · 1.23 KB

Roadmap

Deferred enhancements, captured so the boundary stays intentional. Not started; needs design before code.

Multimodal — remaining gaps

Image input is implemented — user messages on both surfaces, and images inside tool results, which are re-emitted as a follow-up user turn because a probe (2026-07-02) showed upstream silently drops media parts in tool-result outputs (the same workaround the Command Code CLI ships; see D14 for the probe findings and mapping). Still deferred:

  • Document/PDF input. The upstream schema's part union advertises a document block (source-shaped, like Anthropic's); unprobed. Candidate for PDF support on both surfaces later.
  • Vision capability discovery. The model catalog carries no modality metadata; deepseek v4 silently ignores images, GLM-5.2 rejects them, Qwen 3.7 reads them. Callers find out from model behavior. A curated capability map was considered and skipped — it would rot as the catalog changes.
  • Model-side limits. Upstream enforces per-model image constraints (e.g. both dimensions > 10 px on the Qwen backend); the proxy does not pre-validate, it relays the upstream error.