Zero-cost playground¶

Goal: run the GUI against a real Bernstein server with realistic task / agent data, without spending API tokens.

Why¶

The mock adapter (src/bernstein/adapters/mock.py) supports an idle mode that sleeps each spawned agent for a randomized interval instead of calling any LLM. Combined with a sibling fixture repo, this gives you a continuous stream of mock tasks and agents that keep the GUI populated.

Layout¶

<your workspace>/
├── bernstein/             ← this repo (source checkout)
└── bernstein-playground/   ← sibling fixture directory (mock state lives here)

The fixture directory is any empty directory next to your checkout: mkdir ../bernstein-playground is enough. bernstein init inside it creates the runtime state on first run.

The playground is a separate working tree so the mock orchestration writes to its own .sdd/ and never pollutes the source repo's runtime state.

Dev loop¶

Two terminals.

Terminal 1 - mock orchestration¶

cd ../bernstein-playground
bernstein run --idle

Effect:

Forces every agent spawn through the mock adapter.
Sets BERNSTEIN_MOCK_IDLE=1. Each mock agent sleeps BERNSTEIN_MOCK_IDLE_MIN_S..MAX_S seconds (default 15–120) and exits, then the orchestrator spawns a fresh one.
Mutually exclusive with --dry-run. Defined in src/bernstein/cli/run_bootstrap.py (--idle option, lines ~842–1030).
Cost: zero. No LLM calls.

Terminal 2 - Vite dev server with HMR¶

cd <your checkout>/bernstein/web
npm install        # first time
npm run dev

Vite serves the SPA on http://127.0.0.1:5173 and proxies /api/* to FastAPI on :8052.

If you'd rather hit the built bundle (no HMR), run bernstein gui serve --dev in this terminal instead and open http://127.0.0.1:8052/ui/.

Tuning the idle interval¶

Variable	Default	Effect
`BERNSTEIN_MOCK_IDLE`	unset	Set to `1` by `--idle`. Forces idle path.
`BERNSTEIN_MOCK_IDLE_MIN_S`	`15`	Lower bound of per-spawn sleep, in seconds.
`BERNSTEIN_MOCK_IDLE_MAX_S`	`120`	Upper bound of per-spawn sleep, in seconds.

For a denser task stream while iterating on the queue UI:

BERNSTEIN_MOCK_IDLE_MIN_S=2 BERNSTEIN_MOCK_IDLE_MAX_S=8 bernstein run --idle

What you get¶

Tasks page populated with rotating running / queued / done rows.
Agents page shows live token meters from the mock adapter's synthetic counters.
Approvals page exercises the queue + diff layout (mock approval requests if the playground seed includes them).
Costs page reads zeros - mock adapter records no spend. To exercise the cost layout, switch one task off --idle and run a real adapter briefly.