🧠 Responses & Provider Compatibility

Sharper Responses fallback — OpenAI-compatible providers that reject unsupported Responses payloads now retry through Chat Completions without treating every bad request as an unsupported endpoint.
Normalized tool schemas — Function parameter schemas now include an explicit properties object before Responses requests are sent, keeping permissive prompt and MCP schemas compatible with stricter validators.
Plain Responses completions render cleanly — Plain-text final answers in Responses mode now finish the active log entry instead of leaving the WebUI stuck on the generating state.

🔌 MCP Images & Tool Calls

MCP image artifacts — MCP image and image-resource payloads are saved as scoped artifact files, so tool text and attachment metadata can return real file paths while images remain visible to the model.
Streamed tool JSON extraction improved — The parser now prefers complete roots that normalize as tool requests, reducing false misformat warnings when streamed output includes leading prose or incidental JSON.
JSON-string parallel tool calls accepted — Providers that stringify tool_calls inside tool_args are decoded and validated through the normal parallel-tool path.

Message replay ordering fixed — Full log snapshots are treated as authoritative replays, clearing and rendering messages in backend order when sync starts from log 0.
Large backups complete reliably — Backup creation and restore cleanup no longer stop file pattern scans at 50,000 entries, while previews and dry runs stay bounded for responsiveness.

Friendlier skills commands — Empty skills_tool calls now default to listing skills, and the legacy method parameter still works as a deprecated alias for action.
Actionable skill frontmatter warnings — Malformed SKILL.md frontmatter now reports a once-per-path warning with line numbers instead of silently skipping the skill.