From dotfiles to declarative desktops · 5/26 5 min read

Future roadmap: what's next

The February 2026 plan for the dotfiles: a gVisor agent runner, Firecracker microVMs, skill cleanup, and a home server, with a September status update.

Update, 2026-09-30. I wrote this roadmap in February. Since then the new machine turned out to be a MacBook (Parts 6 to 8), and most of the agent runner below is still unbuilt. The status notes in each phase say what happened. The short version: the design documents exist, the Bedrock credential role and a Lima VM for agents exist (Part 15), and the gVisor sandbox, the MCP broker, and Firecracker are not started. I left the original plan below so the comparison is visible.

The goal is a development environment that rebuilds itself from source and runs unattended agents safely. It has five pieces:

  1. A workstation that reproduces itself with git clone and nixos-rebuild switch.
  2. Autonomous agents for ops triage, code review, and routine maintenance.
  3. Agent runners isolated in Firecracker microVMs with no secrets.
  4. A home server for storage, backups, CI, and agent orchestration.
  5. Everything declared in Nix, with each change version-controlled.

Phase 1: the autonomous agent runner

Part 4 explains the design: the agent holds no secrets, and a host-side broker exposes a few audited tools. This phase turns that design into milestones. The design (autonomous-agents-design.md) and an implementation plan covering eight milestones (0 to 7) are written. I condensed them to seven below.

Milestone Scope Status on 2026-09-30
0. gVisor runtime Install runsc, register it as a Docker runtime, run hello-world under it Not started
1. Isolation boundary Minimal sandbox image; confirm ~/.aws and ~/.ssh are unreachable and arbitrary egress fails Not started
2. MCP broker TypeScript MCP server with 7 audited tools (read-only DB, ticket tracker, open-PR) and a full audit log Not started
3. Egress allowlist Host-side proxy that allows Bedrock, package registries, and git hosts and denies the rest Not started
4. Bedrock credential minter Scoped AssumeRole session, injected at job start, 12 hours maximum Partly done: create-bedrock-role.sh and a launchd agent that refreshes credentials exist, but nothing injects them into a sandbox
5. Supervisor systemd --user unit that starts and tears down per-job sandboxes, manual trigger only Dropped. The machine is now macOS, so a launchd-based equivalent would replace it
6. Security audit External review and penetration test of the isolation boundary Not started

Phase 2: Firecracker microVMs

gVisor comes first because it needs no kernel management. The eventual target is Firecracker, which gives each agent its own guest kernel, a dedicated vCPU, pinned memory, and KVM hardware isolation.

In February I described a Firecracker configuration for the Hermes agent under nix/vms/hermes/, with a TAP network, a systemd unit, and a deploy script. It never ran on real hardware. Status: that directory no longer exists in the repo, and Firecracker needs KVM, which macOS does not have. The repo now holds a handoff document (nix/autonomous-agents-firecracker-handoff.md) that marks Firecracker as future work, and the agents that do run today live in Lima VMs. I consider this phase dropped for the Mac and open for the Linux desktop.

Phase 3: Heracles

Update, 2026-10-15: I now do ticket and queue triage through a retro command instead of an ops agent, so this agent was not built.

Hermes is the agent I planned to use for ops triage. Heracles was the name for a larger platform built around it, with four parts: a web UI for task management and audit logs (Studio), scheduled runs (Cron), a registry of MCP tools, and token usage tracking. Status: none of it is built. The repo has a Hermes VM definition in lima/hermes.yaml but nothing named Heracles.

Phase 4: full NixOS switch

The Linux desktop runs Fedora with home-manager on top (home-manager switch --flake .#usman). A full NixOS switch is blocked by an NVMe encryption hardware limitation. nixosConfigurations.default is ready in the flake and waits for that to change.

Status: the new machine I expected to be NixOS was a MacBook, so the macOS port took priority. The Linux switch stays on the list.

Phase 5: skill architecture

Skills were deployed three ways at once: the Nix store, a live git clone, and the npm skills CLI. The target was one mechanism, declared dependencies between skills, and every skill pinned to a commit with hash verification, driven by a CLI:

dotfilesctl skills add <org>/<repo>@<version>
dotfilesctl skills sync
dotfilesctl skills status

Status: partly done, without dotfilesctl. Part 13 covers the consolidation, and the repo now has skills/catalog.yaml and scripts/pin-skill-hashes.sh for pinning.

Phase 6: home server

The old laptop, which runs a few basic services, should become a restic and rclone backup target, a Nix build worker, an agent runner, an Ollama server for local inference, and a Tailscale exit node. It would run the same flake with its own hostname and hardware-configuration.nix. Status: not started. Tailscale now runs on the Mac instead (Part 15).

Phase 7: devshell documentation

The repo had nine devshells, and some were undocumented. The goal was a short README or doc comment on each. Status: there are now ten (browser-use.nix was added), and the READMEs are not written.

Phase 8: dependency audit

The home.packages list was 125 entries on Linux and 88 on macOS in February. I planned to review it quarterly and cut anything untouched for 90 days. Status: I have not counted since. Part 7 includes the audit checklist that covers this.

The schedule as I wrote it

This is a hobby project without deadlines. In February I ordered the work like this:

  • Q3 2026: milestones 0 to 3 of the agent runner.
  • Q4 2026: milestones 4 to 6 and the move to Firecracker.
  • 2027: the full NixOS switch, Heracles Studio, and the skill overhaul.
  • Later: the home server, devshell docs, and dependency reduction.

Q3 is over and milestones 0 to 3 have not started. I no longer commit to these dates.