Skip to content

ROADMAP.md

Latest commit

 

History

History
23 lines (19 loc) · 1.15 KB

ROADMAP.md

File metadata and controls

23 lines (19 loc) · 1.15 KB
Raw

Roadmap

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

Multimodal — remaining gaps

Image input is implemented — Chat Completions user messages, and images inside tool messages, 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; unprobed. Candidate for PDF support 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.