[Feature]: Fully unattended tool invocation by agents

[Power-Maverick]

Provide an idea you would like to see in the tool?

Fully Unattended Tool Execution (Phase 2)

• New optional headless entry point separate from a tool's UI bundle — tool authors ship an agent execution module distinct from their renderer code. Probably need to add a event handler to the PPTB API.
• New manifest contract: pptb.config.json gains an agent section declaring headless action(s), each with its own input/output schema
• Runs in a sandboxed execution context, not the main process directly — third-party code needs the same isolation guarantee BrowserViews already provide today (likely Electron utilityProcess or equivalent)
• No window, no human click — it needs a higher trust bar
• Needs maturity model and only matured tools should be exposed to the agents
• Needs a progress-reporting channel for long-running actions (e.g. reportProgress), since there's no UI to show status and no terminal returnData until the job ends
• Needs explicit per-action risk classification from the tool author (e.g. read-only vs destructive) so PPTB/the agent host can decide what still needs a confirmation step even without a window
• Connection/environment scoping must be passed explicitly per call (no caller window to inherit context from) and checked against a user-configured allowlist before execution. Need more thoughts on auth.
• Needs its own audit/action log in PPTB, since there's no visible window for the user to watch in real time
• Cancellation/timeout handling required, since nothing is bounded by a human closing a window

Relevant log output