Support any OpenAI-compatible LLM provider for scan/generate

[skokaina]

Problem

agentspec scan and agentspec generate are hardcoded to the Anthropic SDK (@anthropic-ai/sdk). Users must have ANTHROPIC_API_KEY to use the two most impressive features. This creates provider lock-in that hurts adoption — users on OpenRouter, Groq, Ollama, or Nvidia NIM can't use core features.

Blocks both launch hooks — neither GIF can be recorded without an Anthropic key, and the launch narrative shouldn't require a specific paid provider.

Proposed solution

Abstract the LLM call behind an LLMClient interface with two backends:

Env vars set	Backend	Providers covered
AGENTSPEC_LLM_API_KEY	OpenAI SDK	OpenRouter, Groq, Together, Ollama, Nvidia NIM, any OpenAI-compatible
ANTHROPIC_API_KEY	Anthropic SDK (existing)	Anthropic, OpenRouter (Anthropic format)
Both set	AGENTSPEC_LLM_* takes precedence	—
Neither set	Error with multi-provider help message	—

New env vars

• AGENTSPEC_LLM_API_KEY — API key
• AGENTSPEC_LLM_BASE_URL — endpoint (e.g., https://openrouter.ai/api/v1)
• AGENTSPEC_LLM_MODEL — model ID (e.g., qwen/qwen3-235b-a22b)

Files to modify

• packages/adapter-claude/src/index.ts — add LLMClient interface, OpenAI backend, env resolver
• packages/adapter-claude/package.json — add openai dependency
• packages/adapter-claude/src/__tests__/claude-adapter.test.ts — tests for OpenAI backend + env resolution
• packages/cli/src/commands/generate.ts — update error message
• packages/cli/src/commands/scan.ts — update error message

Error message when no key is set

No LLM API key configured. Set one of:
  AGENTSPEC_LLM_API_KEY + AGENTSPEC_LLM_BASE_URL  (any OpenAI-compatible provider)
  ANTHROPIC_API_KEY                                (Anthropic Claude)

Free options: OpenRouter (openrouter.ai), Groq (console.groq.com)

Acceptance criteria

• AGENTSPEC_LLM_API_KEY=... AGENTSPEC_LLM_BASE_URL=https://openrouter.ai/api/v1 agentspec generate agent.yaml --framework langgraph works
• ANTHROPIC_API_KEY=... agentspec generate agent.yaml --framework langgraph still works (backward compat)
• No env vars set → helpful error listing both options
• Streaming progress works with both backends
• All adapter-claude tests pass

[iliassjabali]

Closing candidate: PR #17 implements this by extracting codegen into a provider-agnostic @agentspec/codegen package with three backends (ClaudeSubscriptionProvider, AnthropicApiProvider, CodexProvider), auto-resolution via resolveProvider(), a --provider override flag, and an agentspec provider-status command. Streaming, tests, and backward compat through a deprecated adapter-claude shim are all covered.

The only deviation from this issue's spec is the env var naming: PR #17 uses OPENAI_API_KEY / ANTHROPIC_API_KEY directly rather than AGENTSPEC_LLM_API_KEY / AGENTSPEC_LLM_BASE_URL / AGENTSPEC_LLM_MODEL. Happy to add the AGENTSPEC_LLM_* names as aliases in the CodexProvider resolver if we want to match this issue exactly. Otherwise this can be closed once #17 lands.

[iliassjabali]

@skokaina what do you think? Should we add the AGENTSPEC_LLM_* aliases to match this issue's spec, or close it out once #17 merges and stick with the standard OPENAI_API_KEY / ANTHROPIC_API_KEY names?