v2.2.1

CLI: onetool direct

Agent harnesses can invoke tools via subprocess or HTTP rather than MCP. onetool direct is the stable shell contract for that pattern — same tool calls, zero MCP overhead, structured JSON output.

• direct run — execute any tool call from the shell or a script; routes to a running host automatically, falls back to in-process; accepts stdin or a .py file
• direct repl — interactive REPL with tab completion, history, and persistent pack state across lines
• direct start / stop / restart / status / logs — manage a persistent HTTP execution host as a daemon; packs load once and state survives across calls; start blocks until the host is ready
• direct list / search / help — discover and inspect tools without opening a session

Whiteboard

• Persistent sessions — canvas state is now backed by a file; your whiteboard survives server restarts and browser reconnections without losing work
• Live updates — incremental changes are pushed to the browser as they happen, rather than requiring a full page refresh

Tool Improvements

• brave — sources_only=True returns just the source URLs without full result bodies
• webfetch — format='html' returns raw HTML instead of extracted text, useful for scraping structured markup
• tavily — max_sources cap controls how many sources are fetched in deep research mode
• ctx.slice() — integer offsets now accepted in addition to line ranges

Configuration

• Env var overrides — OT_LOG_LEVEL, OT_LOG_DIR, and OT_COMPACT_MAX_LENGTH can now be set without touching onetool.yaml; log lines include full timestamps

Deployment

• Docker — official Dockerfile for containerised deployments

---