Zcode Images

🌍 View in other languages

ZCode image input (explicit native protocol opt-in)

The ZCode executor forwards OpenAI user image_url parts as native image blocks, preserving their position among text parts. PNG, JPEG, WebP and GIF are supported. Data URLs retain their original bytes. Public HTTP(S) URLs use the existing bounded image resolver with DNS/private-address and redirect checks. Limits are 12 images per request and 1 MiB per decoded image. Non-image media is rejected.

The tested official ZCode worker already handles image blocks in its model SDK, but its workspace/generateText wire schema only accepts string content. Enabling images therefore requires an explicitly selected isolated protocol patch. OmniRoute never patches an installed ZCode runtime automatically.

Prepare a separate copy from the supported installed worker:

node scripts/ops/zcode-image-runtime.mjs /path/to/installed/zcode.cjs /path/to/isolated/zcode-images

The helper pins the tested input SHA-256, changes exactly one user-content schema, copies adjacent runtime packages, and records source/output hashes. Unknown builds, duplicate schema matches and conflicting output directories fail closed. Original ZCode files, authentication, provider configuration and native tool behavior remain unchanged. No vendor bundle is included in this repository.

Since 2026-10-18 the helper applies a second versioned step, numeric-provider-code-v1: the finish-record business-error detector accepts only numeric provider codes, so GLM's own generated tool-call text inside a finish record is no longer misread as 502 Provider returned a business error. Explicit failure markers (success:false, event:"error", ProviderBusinessError instances) still fail unchanged. The receipt in omniroute-image-patch.json lists both steps; re-provisioning from a new vendor build requires re-reviewing every step before it applies.

Configure only the OmniRoute process to select the copy through ZCode's command override; retain the existing pristine ZCODE_SERVER_ENTRY and signed-in profile:

ZCODE_AGENT_SERVER_COMMAND=/path/to/node
ZCODE_AGENT_SERVER_ARGS_JSON='["/path/to/isolated/zcode-images/zcode.cjs","app-server","--stdio"]'

Restart OmniRoute through the normal deployment path. Enable image in the caller's model input capabilities for zc/glm-5.3-flash, then verify a real image through that caller. Ordinary zc/glm-5.3 is text-only: the adapter rejects image requests before native inference rather than silently losing pixels. See the official GLM-5.3 documentation. Native inference still returns tool calls to the caller; no autonomous ZCode coding session is created. SSE remains buffered per model step.

To roll back, restore the previous OmniRoute artifact/config, remove the two command overrides, and restart through the same control plane. The pristine installed ZCode worker remains available. Re-review the protocol extension when ZCode is updated; do not apply the patch blindly to a new vendor build.