The GPU budget

One RTX 3070 Ti, 8 GB. qwen3:8b takes ~5.6 GB loaded; nomic-embed-text is small. There is no second slot — a vision model would evict qwen on every swap, which is why the UI loop is deterministic instead.

Standing priority order (ratified by Scott, 2026-09-02): when the model is busy, later entries wait for earlier ones; nothing here ever queues a product feature behind dev tooling.

1. Floor assistant and product features — the agentic MES is the point of the project. Whatever the product ships that needs the local model (assistant routing and answers, instruction drafting, future agent features) goes first.
2. Scored-run log triage — reading each run's logs for what nobody wrote a test for. Bounded per run, and its findings are promoted to store columns.
3. Nightly rollup narration — one note a day; the model is handed computed numbers and asked only what changed.
4. Design chat — Scott's capture surface. Useful, interruptible, and the deep thinking happens in Claude Code anyway.

The UI conformance loop consumes zero GPU by design. Its checks are a real browser and file comparisons. Its entire claim on the model is the one paragraph the nightly rollup already writes, and only when drift exists.

The cloud budget

The floor agent (the Assistant panel's act mode) runs on Claude Sonnet 5 over the API, not the GPU. Cap: $10 a month (Scott, 2026-09-03), set in the Anthropic Console and enforced by the plant: every turn's token usage is appended to ~/.local/share/fsmes/agent-usage.jsonl, the month is summed at list prices, and past MES_AGENT_MONTHLY_USD (default 10) the panel says the budget is spent and the local assistant carries on. A 10-turn floor task with the tool catalogue cached costs roughly 1–3 cents. The Local AI panel shows the running total beside the GPU rows.

Two agents, one bill

There are two agent kinds now — the floor assistant, which proposes changes, and the analysis agent, which holds every read tool and no write tool.

The month's cap is shared. Both kinds' turns go into the same usage file and are summed together, so MES_AGENT_MONTHLY_USD is the whole box's budget whichever agent spent it. A second kind that could get around the month's cap would not be a cap.

A conversation has a cap of its own, and that is the new number:

Key	Default	What it is
MES_AGENT_CONVERSATION_USD	0 (uncapped)	What one floor conversation may spend. Uncapped is what it has always been: a floor conversation is bounded by [admin] agent_max_rounds and by somebody standing at a machine waiting for it
MES_ANALYSIS_CONVERSATION_USD	0.25	What one analysis conversation may spend — a fortieth of the month
MES_AGENT_BRAIN	auto	off switches the floor assistant off
MES_ANALYSIS_BRAIN	auto	off switches the analysis agent off, on its own

Why the analysis agent gets one and the floor assistant does not: it holds every read tool, it is asked to explore, and nobody is watching each round of it. The cap is a speed bump, not a wall — the next conversation starts at zero, and the wall is the month's cap, which no new conversation gets around. When a conversation reaches it, it says so, says what the month has left, and stops; it does not hand over to the local model, because there is no local analysis agent.

These four are environment keys today and become [ai] settings in M2 (docs/design/agentic-harness.md §9), with the same words in the same order: analysis_brain, analysis_conversation_usd. The rename is a move, not a redesign, which is why they are named this way now.

In shadow mode the analysis agent is off and says so (answer 9 of the harness design). The floor assistant falls back to the local model on a shadow plant; the analysis agent has no local stand-in, and answering worse was the option that was turned down.

Embeddings (nomic-embed-text) are effectively free beside qwen and are not budgeted. The Local AI panel on /dashboard/ops displays this order so "is the GPU doing what I decided it should" is answerable at a glance; change the order there and here together, deliberately.