v0.6.6

BasedHardware/omiv0.6.6May 12, 2026by jaberjaber23

AI Summary

Adds support for the latest vLLM version and Discord image handling while introducing new tools and dashboard UI improvements.

Key Highlights

  • vLLM 0.19+ support for the new `reasoning` field.
  • Discord image attachments surface to text-only providers as captions.
  • Channel ID binding hardening to prevent routing issues.
  • New `create_directory` tool with path traversal protection.

New Features

  • vLLM 0.19+ reasoning field support.
  • Discord image attachment handling as captions.
  • New `create_directory` tool.
  • LaTeX math rendering in chat messages via KaTeX bundle.

Full Release Notes

Resolves **11 community issues** and **7 PRs merged** across drivers, channels, and dashboard. 2625 tests passing, zero clippy warnings, zero regressions.

### LLM Drivers
- **#1157** vLLM 0.19+ `reasoning` field support. The driver now reads both `reasoning` (new) and `reasoning_content` (legacy) and emits both on the wire, keeping agentic loops working across vLLM upgrades and other OpenAI-compatible servers.

### Channels
- **#1142** PR **#1143** Discord image attachments surface to text-only providers as a caption + image multipart, so models without vision still get the caption text instead of seeing the image silently dropped.
- **#1144** PR **#1147** channel_id binding hardening. Single source of truth for routing, adapter allowlist, `deny_unknown_fields` on AgentBinding and BindingMatchRule. 14 new tests.
- **#1141** PR **#1176** Shift+Enter inserts a newline in the chat input. Plain Enter still sends.

### Tools
- **#1134** New `create_directory` tool. Routes through workspace sandbox with path traversal protection. 13 aliases (mkdir, make_directory, create_folder, etc.) wired in tool_compat.

### Web Dashboard
- **#1167** PR **#1168** LaTeX math now renders in chat messages via the lazy-loaded KaTeX bundle. CSP updated for jsdelivr.
- **#1140** Chat tab now lists agents from `~/.openfang/agents/`. Boot scans the directory after DB restore and registers any `agent.toml` the DB didn't know about. Bad TOML logged and skipped, never aborts boot.

### Investigations and Design
- **#1097** Workspaces unintended behavior documented (multi-component refactor scope, design comment posted).
- **#1149** OpenAI Responses API migration design-only (affects 26 providers, separate driver gated by flag recommended).

### Quick closes
- **#1048** MiniMax setup answered.
- **#1152** Update command pointed to openfang.sh/update.
- **#1153** Chinese localization dedupe to **#1186**.
- **#1155** 0.0.0.0 bind documented in v0.6.5, comment added.
- **#1156** Agent-as-tool already supported via agent_send.
- **#1159** Server install pointed to docs.
- **#1161** Website restored.

### Stats
- 2625 tests passing (+40 from v0.6.5)
- Zero clippy warnings (`--workspace --all-targets`)
- Zero regressions

**Full Changelog**: https://github.com/RightNow-AI/openfang/compare/v0.6.5...v0.6.6