LLM authentication, provider routing, and tier management

Failure modes

Symptom	Cause	Fix	Severity
`ValueError` from `get_model`	Provider other than `"anthropic"` passed	Pass `"anthropic"`; this provider set is single today	high
`get_model(...)` returns `None`	No model registered for that provider/tier	Check the tier value (`ModelTier.CHEAP.value`, not the enum) and guard the result	high
`ValueError` from `get_tasks_for_tier`	Unknown tier argument	Pass a real `ModelTier`, not a raw string	medium
A task routes to `CAPABLE` unexpectedly	The task string is unknown, so it defaulted	Call `is_known_task` first; add the task to the tier map if it should be classified	medium
`AUTO` returns `API` on a PRO account regardless of module size	`get_recommended_mode` resolves by tier first; `PRO`/`API_ONLY` always return `API`	Expected — only `MAX`/`ENTERPRISE` tiers use size-based selection	medium
Auth always picks the mode you set, ignoring `AUTO` logic	`default_mode` is `SUBSCRIPTION` or `API`, not `AUTO`	Set `default_mode = AuthMode.AUTO` so it defers to `get_recommended_mode`	medium
`await` error calling an executor's `run`	`LLMExecutor.run` is async and was called without `await`	`await executor.run(...)` or drive it with `asyncio.run`	medium
`AdaptiveModelRouter` returns the same model regardless of cost	Fewer than `MIN_SAMPLE_SIZE` calls recorded	Accumulate telemetry; until then it falls back to static defaults	low

Tier value vs enum. Registry lookups take the tier value (tier.value), while routing functions accept either a ModelTier or its string. Mixing the two silently returns None from get_model. Pass tier.value to get_model.
Unknown tasks default silently. get_tier_for_task never raises; an unrecognized task quietly becomes CAPABLE. Validate with is_known_task when the task source is untrusted.
Cost fields are per-million on ModelInfo, per-1k on the properties. input_cost_per_million is the stored field; cost_per_1k_input is the derived property. Don't mix the units when computing an estimate.
get_provider_config is a lazy global. It loads once and caches; call reset_provider_config() after writing a new config file if you need the change to take effect in a long-lived process.

Print the resolved tier: get_tier_for_task(task) — confirms the classification before any model lookup.
Print the model: get_model("anthropic", tier.value) — None means a registry gap or a tier passed as an enum instead of .value.
Inspect auth: attune auth status --json — confirms setup_completed and the active mode.
Inspect provider: attune provider show — confirms the provider and mode the lookups will use.
For routing surprises, read AdaptiveModelRouter.get_routing_stats( workflow, stage) to see sample size and per-model performance.