A desktop-pet end-side agent system with a foreground fast path, sandbox-based verifier, update-first skill lifecycle, and a long-sequence short-task evaluation benchmark.
Layout in two sentences. Everything in
src/,tests/,docs/,scripts/is the project. Every runtime artefact (model weights, caches, skills, sandboxes, benchmark runs, Playwright browser) lives under~/.minideskagent/so thatbash uninstall.shcan wipe it cleanly without touching the rest of your system.
# 1. One-shot install (creates .venv, downloads model, sets up Playwright)
bash install.sh
# 2. Start everything (local model + SSH tunnel + orchestrator + dashboard + pet)
make start
# 3. Run the LongSeqBench (E0/E3/E4/E6 ablations, from_scratch + warmup)
make bench
# 4. Serve the documentation site (architecture + benchmark + runbook)
make docs-serve # http://127.0.0.1:8000
# 5. Wipe everything cleanly
bash uninstall.shsrc/minideskagent/— package source (orchestrator, foreground/background paths, tools, cache, sandbox, verifier, skill lifecycle, benchmark)scripts/— bash entrypoints for the local model server, SSH tunnel, start/stop orchestration, and the benchmark runnerdocs/— mkdocs source (architecture, components, benchmark, results, runbook)tests/— unit / integration / e2e tests
See docs/architecture/overview.md and the rendered diagram
MiniDeskAgent_architecture_final_desktop.svg. The full design document is
MiniDeskAgent_architecture_plan_final_desktop.md.
Every local change is itemised in docs/runbook/local_env_changes.md. The
short version: a project-local .venv, model weights under
~/.minideskagent/models, a Playwright Chromium under
~/.minideskagent/playwright, two background processes (local model +
SSH tunnel) with PID files under ~/.minideskagent/runtime/pids. Nothing is
written outside ~/.minideskagent/ and <project>/.venv and <project>/site.