🧑‍💻 Agent Editor & Scoped Profiles

Build agents from the WebUI — The new Agent Editor offers Easy and Advanced modes for identity, prompts, model presets, tool policies, skill visibility, and avatars without destructive overwrites or model calls.
Choose global or project scope — Sparse profile reads, writes, resets, and policies stay bound to the selected layer, while missing values inherit and explicit empty values clear inherited settings.
Manage the complete profile lifecycle — Create, duplicate, restore, switch, and inspect profiles with scope-aware availability, collision protection, confirmation checks, and running-profile guards.
Create profiles from chat or the Connector — The /profile command and Connector API can create a profile in the selected scope and open a fresh chat with it; /permissions opens the current profile directly in Agent Editor.

🛡️ Tool & Skill Policies

Control tools consistently across every path — A project- and profile-aware resolver applies local, plugin, and MCP tool policies across prompts, schemas, connectors, execution, and delegated agents.
Layered restrictions remain effective — Empty upper-level configs no longer hide inherited limits, canonical identities survive local and MCP execution, and blocked tools disappear from rendered prompts.
Set sparse skill visibility rules — Profiles can allow or block individual skills and define the default for future skills while preserving history and legacy behavior.

Migrate agents from five harnesses — The new Migrate Agents plugin adds a guarded preview-and-import workflow, documentation, WebUI assets, and corrected upload handling.
Blend Kokoro voices natively — The voice plugin expands its American voice catalog with speed controls, validated weighted blending, and reusable in-memory pipelines.
Keep capabilities clearly separated — Slash commands remain in the Commands plugin, skills come only from skills_tool, MCP labels include humanized server names, and capability prompts respect tool availability.
Clean Time Travel history automatically — Background retention sweeps remove orphaned shadow repositories, optionally expire stale history, repair stranded locks, and quarantine corrupt repositories with durable totals and logs.

Document queries block private targets — The fix for CVE-2026-4308 routes remote downloads through the public-only HTTP fetcher and rejects private destinations, including redirects.
Downloads resolve from the filesystem root — Individual and bulk downloads now match the paths exposed by the File Browser and editor.
Malformed chat deletion stops early — Empty or whitespace-only context IDs are rejected before cleanup can resolve them to the shared chat directory.
Plugin edits survive updates — Tracked local changes are stashed during Plugin Hub updates and restored on conflict, with the conflict surfaced in the UI.

Responses fall back on opaque failures — OpenAI-compatible Responses endpoints retry through Chat Completions when they return an opaque 400 error.
Codex OAuth sends the resolved client version — The required version header now satisfies upstream model gates.
Provider history starts with the user — Orphaned leading assistant messages are omitted for Claude-compatible providers.
Project changes save only affected chats — Project updates and deletion no longer rewrite every loaded conversation.
Long-running surfaces do less idle work — Goal revisions arrive through the shared state monitor instead of recurring polls, and elapsed durations format hours correctly.
Desktop and canvas interactions are smoother — Resize drags no longer leak pointer events into the live desktop iframe, and the canvas tab bar can open existing files without closing every editor tab.