# `Agentic.LLM.Provider.CodingAgentBase`

Macro that synthesizes a catalog-only `Agentic.LLM.Provider`
implementation for an ACP-compatible coding-agent CLI (Cursor,
Gemini CLI, Goose, GitHub Copilot, Kimi, Qwen, …).

These agents all route to one or more of the big frontier model
families internally — most expose Anthropic Claude, OpenAI GPT,
and Google Gemini. We surface them in the Catalog as alternative
pathways for those same canonical model families so the
multi-pathway router can score them alongside Anthropic-direct,
OpenRouter, etc.

## Usage

    defmodule Agentic.LLM.Provider.Cursor do
      use Agentic.LLM.Provider.CodingAgentBase,
        id: :cursor,
        cli_name: "cursor-agent",
        label: "Cursor",
        # Optional — defaults to the frontier coding set below.
        # Each tuple is {provider_local_id, label, tier, ctx_window}
        model_overrides: [...]
    end

## Why a macro

Per-agent modules get a single place each (~5 lines) and the
shared catalog/availability machinery lives here. Adding a new
detected agent is a trivial PR. The model list defaults are
intentionally a small "frontier coding" set — agents that route
exclusively to one family override `model_overrides`.

# `build_models`

```elixir
@spec build_models(atom(), String.t(), [{String.t(), String.t(), atom(), integer()}]) ::
  [
    Agentic.LLM.Model.t()
  ]
```

Builds a `[Model.t()]` from a seed list. Called by the generated
`default_models/0` in each per-agent wrapper. Public so the macro
can emit a call to it without escaping the seed list (which would
mangle 4-tuples through the AST representation).

# `default_seeds`

Default model seeds used by every ACP coding agent that doesn't
override them. Mirrors what `Agentic.LLM.Provider.OpenCode`
declared — the canonical_id mapping in `Canonical` then groups
these with their HTTP siblings.

---

*Consult [api-reference.md](api-reference.md) for complete listing*