ADR 0050 — appgen --model is provider-agnostic (any model, not just local Gemma)

• Status: Accepted (implemented 2026-07-13) — @loom/llm gains a GeminiProvider; the CLI --model flag takes a provider spec (PR).
• Date: 2026-07-13
• Amends: ADR 0042 (AI-native brownfield appgen) — specifically its “optional local Gemma model” framing.
• Builds on: the existing @loom/llm LlmProvider SPI + @loom/reasoning ModelRoute ladder.

Context

ADR 0042 shipped the appgen NL→plan stage with an optional model behind @loom/appgen-model, framed as a local Gemma planner (--model <gguf> + node-llama-cpp). In review that default looked premature: it pinned users to a local model with a multi-GB download and a native dependency, when most would rather point appgen at a model they already use (Anthropic, OpenAI, Gemini, …).

Crucially, the architecture was already right: planChangesModel(intent, appModel, { routes }) takes @loom/reasoning ModelRoutes, each carrying any LlmProvider. Only two things pinned it to Gemma: the CLI --model <gguf> flag (which only built a gemmaRoute) and the framing/docs. So this is not a rewrite — it’s exposing the pluggability that was already there.

Decision

--model takes a provider spec; the planner reasons over whichever model is named.

• --model anthropic:<model> / openai:<model> / gemini:<model> → the matching @loom/llm provider, API key from the provider’s env var (ANTHROPIC_API_KEY / OPENAI_API_KEY / GEMINI_API_KEY), checked upfront with a clear error.
• --model <path>.gguf / gemma:<path> / local:<path> → the local Gemma provider (the optional @loom/appgen-model + node-llama-cpp), exactly as before — now one option among several, not the only one.
• --model-big <spec> names a stronger tier for the cheap-first elevation ladder; any spec form.

To make the three named API providers concrete, @loom/llm gains a GeminiProvider (a thin fetch client, same shape as OpenAiProvider/AnthropicProvider). Any LlmProvider works programmatically; these are wired by name in the CLI for convenience.

The fallback contract is unchanged: a missing key/provider, a too-weak model, or an unparseable plan falls back to the deterministic parser — appgen is never left without a valid ChangeSpec.

Consequences

• Local Gemma is no longer privileged — it’s the offline option, chosen deliberately, not by default. Users bring the model they already pay for. @loom/appgen-model stays optional and lazily loaded (its node-llama-cpp native dep is only needed for the local .gguf path).
• The spec-resolution seam lives in the CLI (routeFromSpec), keeping @loom/appgen/@loom/appgen-model free of provider-selection policy. New providers drop in by adding an LlmProvider to @loom/llm and one case to the resolver.
• No change to determinism/provenance: the plan header + lockfile still record which planner ran; the generated code is identical regardless of model.
• Cost/privacy is now the user’s explicit choice (local vs. a hosted API), rather than an implicit local-only default — which is what made the original framing feel wrong.