Make `models` provider-aware instead of OpenCode-only

[kud]

Problem

ai-conventional-commit is now genuinely provider-agnostic, but the models command is not.

Today ai-conventional-commit models still shells out directly to opencode models, and the timeout fallback picker does the same. That means the model list only represents OpenCode-backed models, even though the CLI also supports provider-specific model namespaces such as:

• claude/*
• anthropic/*
• OpenCode-backed provider/model values

So the current models UX is no longer truthful: it presents an OpenCode inventory as if it were the complete model inventory for the application.

Current behaviour

ModelsCommand is explicitly described as List available models via opencode CLI and wraps opencode models. The timeout picker in src/model/picker.ts also calls opencode models directly.

Direction

Make model discovery provider-aware rather than assuming OpenCode is the universal source of truth.

Possible shape:

• define model-discovery capability per provider/adapter;
• aggregate discoverable models from all configured/available providers;
• preserve static aliases where a provider has a small canonical set (for example Claude CLI aliases such as claude/sonnet, claude/opus, claude/haiku);
• use dynamic discovery where the provider exposes it (for example OpenCode);
• handle providers that cannot enumerate models cleanly without pretending they can;
• make models --interactive --save operate over the provider-aware list;
• use the same discovery abstraction for timeout fallback instead of calling OpenCode directly.

The UI should make provider boundaries visible enough that users understand where each model comes from, while still keeping the canonical stored value as the normal provider/model string.

Docs

Update provider/model documentation so ai-conventional-commit models is described as an application-level model picker rather than an OpenCode wrapper.

Also verify model-default documentation while touching this area: the provider docs currently say there is no built-in model default, while the environment-variable table still shows github-copilot/claude-sonnet-4.6 as the default for AICC_MODEL.

Done when

• ai-conventional-commit models no longer represents only opencode models.
• Interactive model selection includes the supported provider namespaces where discovery is possible.
• Timeout fallback uses the same provider-aware model source.
• Providers without dynamic model enumeration are handled explicitly.
• Documentation accurately reflects the multi-provider architecture.

[kud]

Spike notes from a read-only pass over main — no design decisions taken, since the provider-aware shape has more than one reasonable answer.

Call sites, verified this run. Only two places shell out to opencode models:

• src/index.ts:211 — ModelsCommand, which writes stdout.trim() verbatim (the whole formatted table, not just ids) in the non-interactive path, and regex-filters to provider/model ids only for the --interactive picker.
• src/model/picker.ts:10 — fetchAvailableModels, same exec, regex-filtered.

The issue's provider list is one short. createProvider (src/model/provider.ts:71) has four branches, not three: claude/*, codex/* (CodexCliProvider), anthropic/*, and everything else → OpenCode. Any discovery abstraction has to say what codex/* does too.

Three open decisions, each with more than one defensible answer:

1. Enumeration capability. Only OpenCode can list models. claude, codex and the Anthropic SDK cannot be enumerated locally, so it is either hardcoded alias lists (which go stale silently, and for anthropic/* that list is long and moves) or an explicit "known aliases, not exhaustive" state in the UI. Different answers per provider are defensible.
2. Availability detection. Whether a provider appears at all — binary on PATH, ANTHROPIC_API_KEY present, OpenCode session valid — and whether an unavailable one is hidden or shown disabled with the reason.
3. Output contract. This is the one with a compatibility cost: ai-conventional-commit models today prints OpenCode's own table verbatim, so anything aggregated changes what a script piping it receives. Worth deciding deliberately (ids-only? grouped by provider? a --json shape?) rather than as a side effect.

Done and pushed: the model-default doc contradiction named at the end of the issue is real and fixed on autopilot/8-models-default-doc. DEFAULTS in src/config.ts has no model key, so the env table's AICC_MODEL default of github-copilot/claude-sonnet-4.6 contradicted both the code and the prose two paragraphs above it; the cell now reads —. Docs-only, no PR opened.

[kud]

Second spike pass, read-only again; still no design taken. This one pins down the facts the previous notes left open: what each provider can actually list, and what models prints today.

Output contract is narrower than feared. On opencode 1.18.29, opencode models prints bare provider/model ids, one per line, nothing else: 46 lines on this machine, 46 matching the MODEL_LINE_RE filter, zero rejected. So the "verbatim table" that ModelsCommand writes in the non-interactive path (src/index.ts:211) is already an id-per-line list, and the regex filter is a no-op on current output. An aggregated list that keeps one id per line is shape-compatible with what scripts receive now; grouping headers or a --json mode remain a deliberate call, but the compatibility floor is "ids, one per line".

Per-provider listing surface, checked against the installed binaries and SDK:

• opencode: opencode models [provider] filters by provider id and takes --refresh (models.dev cache) and --verbose. Dynamic, local, the only one with a real inventory.
• claude: no list subcommand. claude --help documents --model as "an alias for the latest model (e.g. 'fable', 'opus', or 'sonnet') or a model's full name". Static alias list only, and note haiku is not in that help text today; check before hardcoding the issue's example list.
• codex: no list subcommand; -m, --model <MODEL> is free text. Not enumerable locally (the experimental app-server was not probed).
• anthropic: the SDK exposes client.models.list() (node_modules/@anthropic-ai/sdk/resources/models.d.ts:20). Dynamic but remote, and only reachable with ANTHROPIC_API_KEY, the same precondition validateModel already enforces at src/model/provider.ts:61. So "cannot enumerate" is true locally and false over the network, which is a fourth shape alongside the three above.

Dead file in the way. src/providers/github-copilot.ts is a two-line stub (export * from '../model.js') with no importer in src/ or test/ (grep ran, matched nothing). Anyone creating a providers/ directory for discovery adapters will trip over it; worth removing in the same change.

State of the earlier docs fix: autopilot/8-models-default-doc is still unmerged on origin with no PR open, as of this run.