The 80/20 router

The 80/20 router is one setting: route = "auto". With it, every routine call tries a local Ollama model first. If the response passes the confidence check, the run stops there — zero API cost. If it doesn’t, the router re-runs the same prompt against a Claude model.

Real numbers on a typical founder workload (routines + interactive vinos-ai sessions):

Configure once

~/.config/vinos/ai.toml sets the default router behavior:

# ~/.config/vinos/ai.toml
[router]
local_model  = "llama3.2"
escalate_to  = "claude-sonnet-4-6"
confidence_min = 0.6      # 0–1; escalate when local self-confidence lower
token_ceiling  = 8000     # escalate for prompts larger than this many tokens

[cost]
daily_cap_dollars = 5.00  # hard stop; log a warning and refuse further Claude calls

Per-routine override in the TOML wins:

[agent]
route = "auto"
model = "llama3.2"                        # local first-attempt
fallback_model = "claude-sonnet-4-6"      # what to escalate to

The escalation policy

1. Run local. Local model returns a response plus a self-reported confidence score (via a stop-token pattern the runtime parses).
2. Check confidence. If >= confidence_min, done.
3. Check size. If the input tokens exceed token_ceiling, skip the local attempt entirely — anything that big will just OOM the local model.
4. Escalate. Re-run with fallback_model. This second attempt is the one that costs money.
5. Log both. The ledger records both runs with a route:auto:local or route:auto:escalated tag so vinos-routine cost can tell you which one you paid for.

Which routes what

The 5 shipped routines are already tuned:

• day-brief → anthropic (needs the reasoning; runs once a day; cheap).
• evening-shutdown → anthropic (short output, worth the polish).
• research-recap → ollama (pure summarization; local is great here).
• inbox-triage → auto (most emails are quick; some drafts need muscle).
• github-review → anthropic (code diffs are where Claude earns its keep).

When to override

Use route = "ollama" when:

• The output is short + summarization-shaped.
• You need latency under 1 second.
• The data is sensitive and mustn’t leave the machine.

Use route = "anthropic" when:

• The task is multi-step reasoning across long context.
• You’re generating something (code, prose) where fidelity matters.
• You’ve measured that the local model is failing the confidence check every time anyway.

Use route = "auto" for everything else. That’s the whole point.

Next: troubleshooting routines.