A running list of features to build into hearth later. Not committed work, just a well-organized menu so good ideas do not get lost.
Every idea here should deepen one of hearth's three guarantees or show one off:
- Agents are sandboxed by default.
- Every agent run is audited.
- The whole system is legible and reproducible from boot.
Anything with a release slot has moved to the roadmap (the staged plan from v1.2 to v2.0). What remains here is unscheduled. Effort and impact are rough gut calls to help sequencing.
- Bind-mount filesystem jail. Hide the wider filesystem behind a bind-mount allow list instead of read-only-but-visible. Effort: medium. Impact: medium.
- Multi-agent scheduling with priority queues. Sandboxed workers pull tasks from a queue, each with its own cgroup CPU and RAM cap. Effort: high. Impact: medium.
- Web dashboard mirroring the TUI. A minimal web view of the boot dashboard, reachable over Tailscale. Effort: medium. Impact: medium.
- Signed and attested images. cosign or sigstore: prove the image you booted matches the exact flake commit. Effort: medium. Impact: medium.
- Close the loop with the mcp-audit project. Wire the MCP audit gate so approvals are produced by a real scan from the separate mcp-audit scanner. Two projects, one story. Effort: medium. Impact: medium.
- Boot attestation. TPM-measured boot plus a dashboard panel that says "this system matches commit abc123." Effort: high. Impact: medium.
- Eval arena. Grow run diff into a local benchmark harness: run your models against your own task suite, score with a local LLM judge, leaderboard in the cockpit. Effort: medium. Impact: medium.
- Local tool server. A set of safe, audited tools agents can call (file read within the allow list, web fetch through the egress proxy), exposed to external clients. Effort: medium. Impact: medium.
- Snapshot and rollback of the agent environment. Capture models, secrets, and working state as a Nix closure you can roll back to. Effort: high. Impact: medium.