Roadmap: harden sandboxing, approvals, and server safety

[Hmbown]

Summary

To feel like a first-class coding agent, deepseek-tui needs stronger safety controls around shell execution, sandbox behavior, approvals, and runtime server exposure.

Scope

• Improve sandbox parity across platforms, especially outside macOS
• Add clearer per-tool and per-MCP approval control where appropriate
• Strengthen audit trails for approvals, denials, auto-approvals, and safety escalations
• Make server safety clearer for real-world use, including guarded or authenticated deployment patterns where feasible
• Tighten docs around trust mode, approval policy, and sandbox expectations

Acceptance Criteria

• Approval behavior is easier to reason about for built-in and MCP tools
• Sandbox behavior is more consistent across supported platforms
• Runtime/API/server safety story is clearly documented
• Safety-related regressions are covered by tests where practical

Starting Points

• docs/MODES.md
• docs/RUNTIME_API.md
• docs/CONFIGURATION.md
• crates/execpolicy/
• crates/tui/src/sandbox/
• crates/mcp/src/lib.rs

[Hmbown]

Closing as stale — these April 11 roadmap items predate the v0.4.x / v0.5.0 work and are being replaced by concrete, actionable issues (#28-#35). Will reopen specific items if/when they're picked up.