@cloudflare/codemode@0.5.2 v1.6.0
inboundemail/inboundSep 11, 2026 → Sep 11, 20260 days between releases
What's New
VLM blocks decode predictions natively; adds PostgreSQL sink and remote dispatch.
Breaking Changes
- VLM blocks default to gpt-6-astra, deprecating older versions
- inference server start now publishes on localhost by default
New Features
- VLM blocks decode predictions natively with shared decoding package
- Native PostgreSQL sink added
- Remote dispatch for inner workflows
- Per-source telemetry for stream pipelines
- Zero-shot action recognition model ID updated to nvidia/cosmos-3-edge-action-recognition
Improvements
- Shared decoding package for VLMs reduces duplication
- Updates action recognition model ID to nvidia/cosmos-3-edge-action-recognition
- Adds GPT-6 Astra support
Bug Fixes
- Answers in wrong contract now set error_status instead of silently returning zero detections
Upgrade Notes
Swap VLM steps to new versions to access predictions directly and delete formatter steps. Update model IDs to gpt-6-astra or nvidia/cosmos-3-edge-action-recognition. Configure localhost publishing if needed.
The Two Releases
@cloudflare/codemode@0.5.2
Sep 11, 2026· 1 month ago
Improved JSON truncation logic to handle oversized objects structurally, preventing invalid JSON output. Added `toModelOutput` projection and updated MCP servers to apply response budgets.
Full release notesv1.6.0
Sep 11, 2026· 1 month ago
VLM blocks now decode their own predictions, adding a Native PostgreSQL sink, remote workflow dispatch, and per-source telemetry. Updates action recognition model ID and adds GPT-6 Astra.
Full release notes