tau-router

Two Tau extensions for spending across several LLM accounts on a budget:

• router picks the model for each request from a profile: an ordered list, best first. It moves down a rung when a provider's budget runs low or a model is rate-limited, and back up when the block clears. A model you pick with /model is pinned until you resume routing.
• planner gives the model a delegate tool. Each part of a job runs as a sub-agent on its own model, with its own budget slice and a shell command that says when it's done. A failed check is retried once, one rung up.

Status: 0.1.0, working toward 1.0 (see ROADMAP.md). It works on Tau 0.13.0. Spend is tracked per process, so two Tau instances don't yet see each other's spending; 0.2 fixes that.

Install and load

pip install ffwf-tau                        # Tau 0.13.0 or later, from PyPI
pip install ffwf-tau-router                 # into the same environment as Tau
tau $(tau-router ext-args)                  # loads router and planner

Tau 0.13 loads extensions only from files, so tau-router ext-args prints a -e <path> for each installed module. Nothing is copied into ~/.tau. The package also declares tau.extensions entry points, which Tau will read once john/tau #35 lands. Then tau alone finds the extensions, and you enable them in config.

/plan <goal> is a prompt template that ships in the package. Tau 0.13 reads templates only from its prompt directories (john/tau #39). Until #39 lands, link it in yourself if you want it:

ln -s "$(tau-router prompt-dir)/plan.md" ~/.tau/prompts/plan.md

delegate works without the template. Asking for delegation in plain words is enough.

Configure

The router reads extensions.router in Tau's config.json:

"router": {
  "profiles": {"budget": ["claude-haiku-5-5", "mistral-small", "nemotron-super"],
               "free-only": ["nemotron-super", "flash-lite"]},
  "profile": "budget",
  "budgets": {"anthropic": {"cost": 3.00, "period": "month", "anchor_day": 8},
              "mistral":   {"cost": 5.00, "period": "month", "anchor_day": 1},
              "cerebras":  {"cost": 15.00, "period": "once"}},
  "currency": "$"
}

• Profile names are model names from models. The router prices each call with that model's cost block, the prices per million tokens.
• A budget is a cap on what this router spends with a provider in each period: day, month from anchor_day (the day of the month it resets, 1–31, in UTC; default 1; in a month too short for it, the last day), or once since since. Your other tools share these accounts, so the cap is an allowance, not a balance.
• cost amounts are in whatever currency your prices are written in. currency is only the display symbol, and nothing is converted.
• A model with no cost block is unpriced. You can pin it by hand, but the router won't route to it under a budget.

Other keys:

Config keys from before 0.1 (usd, persona_usd, reserve_usd) are refused, with a message naming each key's new name.

Commands and tools

In the TUI, the status strip shows router: <model> [provider] · <profile or pinned> <spent>/<cap> <period>.

Tests

pip install -e '.[dev]' && pytest. The tests run the extensions against a fake Tau api and ctx with scripted turns. No model is called. experiments/ holds the live runs, results (FINDINGS.md) and a sandbox for trying it (TEST-DRIVE.md).

License

MIT.