nc://fallback

Fallback: heuristic model failover, ranking your own providers' models and keeping Nodecode's failover chain filled. A Nodecode rewrite of tomkubik/hermes-fallback by tomkubik.

Nodecode already walks a chain of fallback models when a round's own retries are spent on a rate limit, an outage, a broken answer or a refusal. What it does not do is choose the models in that chain: that is a job for whoever is still awake. So this cell ranks every model your configured providers offer against what a turn needs -- a context window big enough, a model big enough, a recent one, a free one, one that can call tools, one on the provider you trust first -- and writes the best few into failover.chain. It makes no model call, so it works while the model you are on is the one that is down.

Get it

Tell the agent "add the fallback cell", or run nodecode add nc://fallback.

Use it

Ask for it, or call it:

• fallback_resolve -- the ranking, without writing anything.
• fallback_apply -- the ranking, written into failover.chain.
• fallback_status -- the primary model, the chain, and the providers this box can name models for.
• /fallback status | resolve | apply | clear | providers [key=value ...], the same from the prompt, when the model that would answer is the one that just failed:

/fallback resolve min_context=128k sort_by=latest free_only=true
/fallback apply count=3
/fallback clear

• From eval: (fallback:resolve), (fallback:apply-chain :max-candidates 3), (fallback:status), (fallback:providers). (help :fallback) is the manual.

Criteria, the section's or a call's: provider (auto, existing, or one id), providers (a preference order, which also limits the search), exclude_providers, sort_by (largest_parameter_count, greatest_context, latest, latest_flash, smallest, provider_preference), min_context/max_context (128k, 1m, 200000), min_params/max_params (30b), free_only, flash_only, require_tools, vendor, name_filter (a substring, or a * glob), prefer_current_provider, allow_provider_switch, max_candidates.

Where the numbers come from, in this order: the model id itself (70b, 128k, :free, 4o-mini), then Nodecode's own models.dev row for the id, then a small table of families whose ids publish none. A number nobody published stays unknown, and a criterion that needs one leaves that model out rather than guessing; require_tools drops only a model whose catalog row says it cannot call tools.

Set it up

Nothing is required. The section is optional, and auto_sync is on by default: while failover.chain is empty, the cell fills it at start from the section's criteria and says so. A chain somebody wrote by hand is never touched -- say /fallback apply to replace it.

{
  "fallback": {
    "sort_by": "largest_parameter_count",
    "min_context": 128000,
    "require_tools": true,
    "max_candidates": 3,
    "auto_sync": true
  }
}

No key, no endpoint and no account: it reads the providers you already configured and the catalog Nodecode already caches.

The credit

MIT, Copyright (c) 2026 Tom Kubik: see LICENSE and NOTICE.